LEARN · SIMULATE · VERIFY

Learn

Tools Blog

Help

Home

This page in other languages:한국어 원문

English

Weekly public experiment · locked before the draw · scored automatically

Can AI Predict Korean Lotto 6/45 Winning Numbers?

Name: NumLM AI lottery experiment — pre-published combinations and scoring by draw
Creator: NumLM

Every week we test whether statistical, LSTM, Transformer and ensemble models beat random picks on Korea's Lotto 6/45, using the full draw history.
Each method publishes 100 combinations before the draw; after the draw they are scored automatically and the results accumulate.

This week — Draw #1248, 100 combinations Cumulative results

At a glance

The NumLM AI lottery experiment is a public, weekly test of whether statistical, LSTM, Transformer and ensemble models trained on the full Korean Lotto 6/45 history beat uniform random picks, judged only by the cumulative scoring of 100 combinations per model published before each draw.

Five arms — uniform random (control), a statistical Markov-chain method, an LSTM, a Transformer decoder and the Global Lottery Ensemble — each publish 100 combinations per draw.
Combinations are generated and published before ticket sales close (Saturday 20:00 KST) with a recorded timestamp; nothing is generated or edited after the cutoff.
Expectation for a perfectly random picker (hypergeometric): 0.8 matches per combination, about 2.38 combinations with 3+ matches per 100, and jackpot odds of 1 in 8,145,060.
The holdout validation loss of the LSTM and Transformer never goes below the random-draw floor ln C(45,6)/7 ≈ 2.273 — a measurement that the models learn the format but not the content.
Seeds are public (draw × 1000 + method id; FNV-1a hash for the ensemble), so anyone with the same history can reproduce the same combinations.

Live measurements (server-rendered, current as of this page load)

Across 9 scored draws, average matches per combination — Random 0.773 · Stats 0.756 · LSTM 0.763 · Transformer 0.842 · Ensemble 0.804 (expected 0.800).
Cumulative combinations with 3+ matches — Random 25 · Stats 18 · LSTM 15 · Transformer 31 · Ensemble 12 (expected 21.45 per method).
Latest scored draw #1247 (2026-10-24): winning numbers 8, 14, 21, 22, 42, 45 + bonus 24; average matches by method Random 0.80 · Stats 0.25 · LSTM 0.72 · Transformer 0.92 · Ensemble 0.70.
Current draw #1248: combinations published 10/24/2026, 20:46 KST; sales cutoff 10/31/2026, 20:00 KST.

Try it yourself: Open the Transformer used here · Markov chain experiment · Open data (JSON)

New to Korean Lotto 6/45? Read this first

Lotto 6/45 is South Korea's national lottery: players pick 6 different numbers from 1 to 45.
One draw every Saturday evening; 6 main numbers plus a bonus number are drawn without replacement.
The odds of matching all 6 numbers with one ticket are 1 in 8,145,060 (C(45,6)). Each number appears in a draw with probability 6/45 ≈ 13.3%.
Before each draw, NumLM publishes 100 combinations from every model on this page and records the timestamp; nothing is changed afterwards.
After the draw, each set of 100 is scored automatically. If no method consistently beats the random control, that is the finding — the page is a verification, not a tips service.

Rules — numbers are published before the draw and never changed afterwards

① Before ticket sales close (draw day 20:00 KST), five methods — Random · Stats · LSTM · Transformer · Global Lottery Ensemble — each generate 100 combinations that are published here in advance (timestamp recorded; nothing can be generated after the cutoff). ② The LSTM and Transformer are genuinely trained on the full draw history, and every seed is derived deterministically from the draw number, so anyone can reproduce the same combinations. ③ Results are scored automatically after the draw and accumulate permanently — there is no way to show only the good weeks. The goal is verification, not prediction. No conclusion is fixed in advance — we check every week whether any method consistently beats the random baseline (the hypergeometric expectation). Why uniform random picks are the control group is explained on the uniform distribution page (Korean).

This week's experiment — Draw #1248 · published before the draw

Published: 10/24/2026, 20:46 KST · sales cutoff: 10/31/2026, 20:00 KST · data: full history through draw #1247 (1247 draws); draws #1–#1227 used for training, the latest 20 draws (#1228–#1247) reserved as a validation holdout · seed rule: draw × 1000 + method id (reproducible) · Markov stats: co-occurrence transitions over the last 199 draws, current state = draw #1247 [8, 14, 21, 22, 42, 45], candidate pool = top 18 (previous-draw numbers stay eligible — 0 of them are in this pool)

What this week's training tells us — holdout validation loss: Transformer 2.6603 · LSTM 2.6444. The theoretical floor for a random draw is 2.2733 (ln C(45,6)/7), so on draws they have never seen neither model even reaches the floor — a measurement of "the models learn the format (sorted numbers, separators) but not the content (which combination)".

Uniform random (control)

Six numbers drawn uniformly from 1–45 with no analysis at all. The question is whether anything beats this.

Statistical — Markov chain

Estimates transition probabilities P(Y|X) from co-occurrence in the last 200 draws and scores each number by the average P(Y|X) over the previous draw's six numbers; those numbers stay eligible (in real draws about 0.8 of them repeat). Weighted sampling from the top 18. An honest implementation of the "numbers that come together" strategy.

LSTM (actually trained)

Autoregressive sampling from an LSTM trained on the draw history as a token stream (next-token prediction).

Transformer decoder (actually trained)

Autoregressive sampling from a causal Transformer decoder trained on the same stream — the same architecture as the Decoder Atlas lab.

Global Lottery Ensemble

New arm

Combines traditional lottery heuristics (Funatsu rules, delta and positional frequency), hot/cold and gap statistics, pair/triple co-occurrence, Bayesian and Markov models, machine learning, LSTM and Transformer predictions into 100 combinations. Experimental — it does not improve your odds. Joined the experiment at draw #1248 · generated 10/24/2026, 20:46 KST · version WE-1.1. Its cumulative record starts from that draw; earlier draws are not back-filled.

Draw #1247 results (2026-10-24)

Winning numbers

(bonus)

Numbers matched	Random	Stats	LSTM	Transformer	Ensemble	Expected (hypergeometric)
0	41	75	45	31	42	40.06
1	43	25	41	48	46	42.41
2	12	0	11	19	12	15.15
3	3	0	3	2	0	2.24
4	1	0	0	0	0	0.14
5 🎯	0	0	0	0	0	2.9e-3
6 🎯	0	0	0	0	0	1.2e-5

Per method, 100 combinations ("—" = the method did not take part in that draw). Expected = the number of combinations a perfectly random picker would produce — if every method sits near it, none of them is doing better than chance.

Random: avg 0.80 · best 4

Stats: avg 0.25 · best 1

LSTM: avg 0.72 · best 3

Transformer: avg 0.92 · best 3

Ensemble: avg 0.70 · best 2

Expected average 0.80 per combination

Cumulative results — the picture sharpens as draws accumulate

9 draws scored · 900 combinations per method (4100 in total) · Global Lottery Ensemble: 5 draws since #1243 (new arm — no back-filling) · the "Stats" column combines frequency weighting for 2 draws (through #1240) + Markov chain excluding previous numbers for 2 draws (from #1241) + Markov chain including previous numbers for 5 draws

Numbers matched	Random	Stats	LSTM	Transformer	Ensemble5 draws · 500 combos	Expected (hypergeometric)
0	381	370	362	353	198	360.51
1	369	399	404	370	214	381.71
2	125	113	119	146	76	136.33
3	23	17	15	28	12	20.20
4	2	1	0	3	0	1.23
5 🎯	0	0	0	0	0	0.03
6 🎯	0	0	0	0	0	1.1e-4

Per method, 900 combinations ("—" = the method did not take part in that draw). Expected = the number of combinations a perfectly random picker would produce — if every method sits near it, none of them is doing better than chance.

Metric	Random	Stats	LSTM	Transformer	Ensemble	Expected
Average matches per combination	0.773	0.756	0.763	0.842	0.804	0.800
3+ matches (cumulative)	25	18	15	31	12 / expected 11.92	21.45
Best match ever	4	4	3	4	3	—
Draws taken part (since)	9 (from #1239)	9 (from #1239)	9 (from #1239)	9 (from #1239)	5 (from #1243)	—

Average matches by draw

Draw	Random	Stats	LSTM	Transformer	Ensemble
#1247	0.80	0.25	0.72	0.92	0.70
#1246	0.90	1.18	0.69	0.85	0.88
#1245	0.80	0.73	0.61	0.69	0.59
#1244	0.76	0.97	0.82	0.77	0.89
#1243	0.60	0.66	0.77	0.87	0.96
#1242	0.83	0.31	0.75	0.98	—
#1241	0.75	1.05	0.98	0.78	—
#1240	0.73	0.85	0.67	0.87	—

Global Lottery Ensemble — does combining many methods beat random?

Combines well-known lottery heuristics with statistical, probabilistic and AI signals into 100 combinations per draw. No conclusion is fixed in advance — only the cumulative scoring of pre-published combinations answers the question. Joined at draw #1243 · earlier draws are not back-filled.

Traditional

Funatsu rules · Hot/Cold · Gap

Relationships

Pair · Triple · Markov

Probability & ML

Bayesian · Machine learning

Deep learning

LSTM · Transformer

Many signals → validated on past draws (walk-forward) → ensemble score → 20,000 candidate combinations → near-duplicates removed → 100 combinations. The weights are not hand-picked: they are learned from predicting the last 20 draws using only earlier history (no single method above 30%).

Walk-forward performance (leave-one-out) — log loss of the winning numbers 3.811 vs uniform 3.807 · mean percentile of winning numbers in the ensemble score 0.493 (no information = 0.500) · top-6/10/15 recall 12% / 17% / 34% (no information = 13% / 22% / 33%). For a random draw these figures merely fluctuate around the no-information baseline.

Number layer — weight per signal (group → within group)

Traditional statistics 20%	Hot / Cold 10% · Gap / Overdue 10%
Relationships 20%	Markov 20%
Probability 20%	Bayesian 20%
Machine learning 19%	Machine learning 19%
Deep learning 21%	LSTM 10% · Transformer 10%

Combination layer — weight per combination property

Number ensemble score 18% · Funatsu Sakai rules 23% · Pair 33% · Triple 12% · Structure 13%

Hot / Cold: Average standardised deviation of each number's count from its expectation (window × 6/45) over the last 10, 24, 50, 100 draws and the full history. Describes the past; it does not change the next draw. Gap / Overdue: Current gap since the last appearance ÷ historical mean gap, clipped to 0–3. A description of the drought, not a "due" signal. Bayesian: Posterior mean of each number's inclusion probability under a Beta prior. The prior is strong (worth ~300 draws) so small frequency differences are not mistaken for "ability". Markov: Conditional frequency P(n_t | j_{t-1}) that n follows j from the previous draw, shrunk towards 6/45 and averaged over the previous draw's six numbers. Machine learning: Logistic regression on draw × number rows (frequency deviations, gap, previous-draw inclusion, co-occurrence with the previous draw, Bayesian and Markov scores, parity, range), trained strictly walk-forward in time order. LSTM: Inclusion frequency of each number in combinations sampled from this experiment's LSTM (actually trained on the draw history). Transformer: The same inclusion frequency from this experiment's Transformer decoder. Funatsu Sakai rules: A weak prior — each number gets a continuous weight from its count in the last 30 draws (3–4: 0.95, 2/5: 0.61, 1/6: 0.25, 0/7: 0.07) and the combination scores 1.0 for exactly one number carried from the previous draw (0 or 2: 0.4), averaged half and half. No number is removed; its share of the combination score is capped at 25%. Pair: Mean Laplace-smoothed co-occurrence lift of the 15 pairs in the combination. Triple: Mean co-occurrence lift of the 20 triples in the combination. Triples are very sparse (0.42 expected appearances each in 300 draws), so they are shrunk toward 1 much harder than pairs (from WE-1.1, about 9% weight on the observed value). Structure: Mean log-likelihood ratio of the combination's odd count, low/high count (1–22 / 23–45), sum bucket and consecutive pairs against the distribution of random combinations. Never used as a hard filter.

Top 10 numbers by ensemble score this draw (0–1, weighted sum of percentile-normalised signals)

27 · 0.76

18 · 0.75

33 · 0.71

12 · 0.70

34 · 0.69

45 · 0.68

15 · 0.67

13 · 0.67

20 · 0.63

40 · 0.62

The 100 combinations are not simply the top numbers bundled together: 20,000 candidates are sampled with softmax(score/T, T=2), scored as combinations, then placed in two steps. ① Band quotas — 90 combinations split across frequency bands from the last 30 draws (core · core+low · core+high · other) in proportion to the chance that a real, uniformly random draw lands in each band, plus 10 random exploration combinations that ignore the number scores. ② Coverage optimisation — jointly maximising total score + covered pairs and numbers − combinations sharing 3+ numbers − concentration on a few numbers (Gail Howard's wheeling idea). Placement does not change the expected number of wins; it only reduces weeks with no 3-match at all. Reproduction: version WE-1.1 · seed 2530565173 (= FNV-1a("1248|WE-1.1|1247")) · history through draw #1247 · generated in 151.4s.

Combination scores (top 5)

#	Combination	Total	Numbers	Funatsu	Pair	Triple	Structure
1	6 12 15 29 40 45	2.06	3.37	0.94	1.24	1.21	-0.04
2	10 13 16 19 21 33	1.89	3.53	0.78	1.23	1.13	0.06
3	1 13 18 24 27 45	2.08	3.93	0.72	1.27	1.09	0.06
4	3 7 22 27 34 38	1.46	3.27	0.80	1.24	1.08	-0.02
5	2 11 17 26 31 43	-0.26	2.75	0.50	0.97	1.03	0.06

Experimental combinations based on past patterns; they do not improve your odds. The arm is compared with the single methods (random, stats, LSTM, Transformer) under identical conditions (100 combinations), and no single draw decides which is better.

Why even AI cannot predict the draw — what this experiment teaches

The models are not dumb. The same Transformer and LSTM learn sequences with structure (text, patterned data) easily — in the verification script the loss on patterned data falls to less than half the floor. A fair draw simply contains no pattern to learn.
"They learn the format, not the content" — the loss does go below ln 45, down to the floor of the sorted encoding, ln C(45,6)/7 ≈ 2.273 (ascending order, separator tokens). On draws the models have never seen it never goes lower. That wall is randomness itself.
The same holds for "hot numbers" (frequency) and "numbers that come together" (Markov transition probabilities) — in a random draw those differences are sampling noise, so with enough draws they become indistinguishable from uniform random picks. The statistical arm used frequency weighting through draw #1240 and the Markov chain technique from #1241.
"Does combining many methods help?" — the Global Lottery Ensemble merges Funatsu rules, hot/cold, gap, pair/triple, Bayesian, Markov, ML, LSTM and Transformer signals with walk-forward-learned weights. If each signal is noise, the combination is noise too — whether that is so is answered by the cumulative table.
The jackpot odds are 1 in 8,145,060 per combination. Even 500 combinations a week for a whole year give an expected 0.0032 jackpots. The cumulative table on this page reproduces that arithmetic with real draws.

Markov transition probabilities →

Open the Transformer used here →Bayesian network experiment Markov chain experiment

This page is for education and entertainment. No method improves your odds, and any lottery purchase is your own responsibility regardless of the numbers shown here.

Find other concepts and labs

개념·랩·가이드 검색

Ctrl K

FAQ — can AI really predict the lottery?

Q. What is Korean Lotto 6/45?

South Korea's national lottery, drawn every Saturday. Players choose 6 different numbers from 1 to 45; 6 main numbers and a bonus number are drawn without replacement. Matching all 6 wins the jackpot at odds of 1 in 8,145,060. Draws are independent, so a fair draw carries no memory of previous results.

Q. Can AI predict Korean lottery numbers?

Not according to the measurements on this page. The LSTM and Transformer are genuinely trained on the full draw history, yet their loss on held-out recent draws never goes below the random-draw floor ln C(45,6)/7 ≈ 2.273, and the scored match distributions stay near the hypergeometric expectation. Results accumulate here every week.

Q. Can past draw data predict future winning numbers?

A fair draw is independent of every previous draw, so history does not change the next draw's probabilities. Under that assumption, the differences in frequency, gaps and co-occurrence seen in past data are better explained by sampling variation than by a persistent signal. This page does not ask you to take that on faith: every week 100 combinations are published before the draw and scored afterwards.

Q. Do LSTM or Transformer models beat random picks?

Compare them yourself in the cumulative table. A purely random picker averages 0.8 matches per combination and about 2.4 combinations with 3+ matches per 100. So far random, statistical, LSTM and Transformer arms all sit near those expectations; any single-draw difference is within chance.

Q. Are all numbers really equally likely?

Yes. Each of the 45 numbers appears in a draw with probability 6/45 ≈ 13.3%, and each of the 8,145,060 six-number combinations is equally likely. Note that "each number is equally likely" is not the same as "six independent uniform draws": the balls are drawn without replacement, so no number repeats within a draw and the six are not independent.

Related: Uniform distribution (Korean)

Q. Is it better to pick frequently drawn numbers?

No. Every six-number combination has the same 1 in 8,145,060 chance. The statistical arm of this experiment is an honest implementation of exactly that strategy (frequency weighting, then Markov co-occurrence), and its cumulative record is indistinguishable from random picks.

Q. What is the Global Lottery Ensemble?

The fifth arm of the experiment. It combines traditional lottery heuristics (Funatsu rules), hot/cold and gap statistics, pair/triple co-occurrence, Bayesian and Markov models, machine learning, LSTM and Transformer predictions into 100 combinations, with weights learned walk-forward on past draws. It tests whether combining many methods beats random; its record starts from the draw it joined. It does not improve your odds.